Most Sanskrit speech datasets are built for automatic speech recognition or conversational text-to-speech, so they do not preserve the strict rhythmic structure of verse. In Chandas, the classical prosodic framework, each syllable is classified as laghu (light) or guru (heavy), and recitation must follow that meter. A new arXiv preprint introduces ShlokaDhara, a 43-hour, multi-speaker corpus designed specifically for text-to-recite: synthesizing spoken Sanskrit verse with both prosodic accuracy and metrical correctness.
ShlokaDhara includes multiple meters, recitation styles, and melodic variations. The authors benchmarked state-of-the-art TTS systems using UTMOS, an automated MOS predictor, and report that XTTS-v2 achieved the highest perceptual quality. They then validated XTTS-v2 outputs with two additional checks: subjective ratings from 16 Sanskrit experts and an objective syllable-level analysis of laghu-guru durations to measure structural metrical awareness.
Because the paper is a single source, there is no independent comparison to weigh here; the claims are the authors' own. Still, the work offers a dataset and evaluation benchmark for prosody-aware computational Sanskrit recitation, filling a gap left by ASR- and conversation-oriented corpora.